Monolithic AI vs Modular AI

[@code4AI]

Please note, with the automatic dubbing from KZbin /Google you hear a synthetic voice in your regional language. To hear my original voice in English, switch to "Default" or "English" in the settings. Thank you.

[@davidtindell950]

thank you for your ENTHUSIASM ! ! ! !😊

[@code4AI]

Always!

[@CharlotteLopez-n3i]

Love the emphasis on modularity and specification in building reliable AI systems! The organic solar cells example is a great illustration of how a model-based approach can lead to more efficient solutions.

[@aennmatyasbarra-hunyor5506]

Great series

[@code4AI]

Thank you.

[@AshWickramasinghe]

Working a system that has the same core concept 😅 we shouldn't need to wait for a few companies to give us access to their expensive monolithic models for us to use GenAI for our daily tasks.

[@I_am_who_I_am_who_I_am]

😅 here is enterprise AI coming 😂😂😂😂 Microservices 😂😂😂😂

[@pensiveintrovert4318]

So their genius idea is that a specialist is better at its specialty than a generalist? I can expect this kind of insight from a chatGPT when I ask it to explain something to a five year old. That is why mixture of experts works better than a monolith model.

[@code4AI]

No. The idea is, that you have to design your AI architecture and the different NN configurations in your layers in a specific way, and that your specific task is the main design criterium for a dedicated and optimized AI architecture.

[@ibgib]

This is highly close to what I've been building: semantic version control. It's like a generalized git. Specification is one layer on top of version control. The protocol enables content addressing (like git commit hashes) to version specs, models, training data, metadata like agentic architectures, and even derivative data. This very instant, I'm working on agent-based JIT UX. Still looking for investment!